Visual Perception of 3D Worlds for Haptic Exploration

我们生活在一个非常视觉化的世界里;我们主要依靠我们的眼睛来告诉我们我们在哪里，我们周围有什么，然后根据感知到的信息决定去哪里以及如何到达那里。甚至交通标志和指示灯也依赖于我们的眼睛。GPS（全球定位传感器）通过为我们提供另一种传感器来告诉我们我们在哪里，从而彻底改变了个人导航。不幸的是，GPS需要事先的地图以及至少3个GPS卫星的清晰视野，以便三角测量我们的位置。如果我们开始就没有环境地图，或者我们不能依靠GPS来告诉我们我们在哪里，那会怎么样？相机是一种传感器，非常类似于我们的眼睛。Visual SLAM（Simultaneous Localization And Mapping）是在没有先验地图的情况下，仅使用摄像机同时构建地图和定位的问题。视觉SLAM是一个相对较新的研究领域，已经显示出很大的希望，但受到鲁棒性和计算易处理性问题的阻碍。当前的方法依赖于指数累积的单个特征点，特别是在真实的世界环境中。我们建议使用真实的世界对象（或特征集群）作为地标，而不是特征点。我们已经证明了特征点的三元组可以导致更好的对象检测和识别。使用这些点或对象本身的三元组可以产生有意义的紧凑SLAM地图并解决当前的缺点。我们还提出了其他改进的视觉SLAM，来自结构从运动（SFM）技术，如过滤或错误建模中使用的技术。

We live in a very visual world; we rely chiefly on our eyes to tell us where we are and what is around us and then make decisions of where to go and how to get there based on that sensed information. Even traffic signs and indicators rely on us using our eyes. GPS (Global Positioning Sensors) have revolutionized personal navigation by providing us with another sensor for telling us where we are. Unfortunately a GPS requires a prior map as well as clear sight of at least 3 GPS satellites in order to triangulate our position. What if we had no map of our environment to begin with or we could not rely on a GPS to tell us where we were? A camera is a sensor that closely resembles our eyes. Visual SLAM (Simultaneous Localization And Mapping) is the problem of building a map and localizing at the same time without a prior map by only using a camera. Visual SLAM is a relatively new research area that has shown great promise but is hindered by the problem of robustness and being computational tractable. Current methods rely on single feature points which accumulate exponentially, especially in real world environments. We are proposing to use real world objects (or clusters of features) as the landmarks as opposed to feature points. We have already demonstrated that triads of feature points can lead to better object detection & recognition. Using these triads of points or the objects themselves can lead to meaningful compact SLAM maps and address current shortcomings. We also propose other improvements to visual SLAM that come from techniques used in Structure From Motion (SFM) techniques such as filtering or error modelling.